A Software Engineering Perspective on Testing Large Language Models: Research, Practice, Tools and Benchmarks

TL;DR

To address safe deployment of LLMs in software systems, the paper advocates organizing LLM testing through a software engineering lens by extending the ML testing taxonomy with LLM-specific topics. It reports a taxonomy-based survey of current LLM research, open-source benchmarks, testing tools, and practitioner discussions to identify coverage and gaps. The study finds that public tooling and benchmarks address only a subset of the taxonomy (primarily correctness) while critical topics such as privacy, security, efficiency, interpretability, and data/learning-program testing are underrepresented. It argues for closer cooperation between SE and ML communities and for taxonomy-driven tooling and benchmarks to enable safer, scalable LLM testing in 2030.

Abstract

Large Language Models (LLMs) are rapidly becoming ubiquitous both as stand-alone tools and as components of current and future software systems. To enable usage of LLMs in the high-stake or safety-critical systems of 2030, they need to undergo rigorous testing. Software Engineering (SE) research on testing Machine Learning (ML) components and ML-based systems has systematically explored many topics such as test input generation and robustness. We believe knowledge about tools, benchmarks, research and practitioner views related to LLM testing needs to be similarly organized. To this end, we present a taxonomy of LLM testing topics and conduct preliminary studies of state of the art and practice approaches to research, open-source tools and benchmarks for LLM testing, mapping results onto this taxonomy. Our goal is to identify gaps requiring more research and engineering effort and inspire a clearer communication between LLM practitioners and the SE research community.

Figures (1)

• Figure 1: Taxonomy of ML testing in SE ZhangHML22 with additional LLM-specific topics.